Nvidia's Physical AI Stack Underpins Skild AI's S1 Robot Model

What happened

Nvidia’s blog post, published September 10, 2026, walks through the technical partnership behind Skild AI’s S1 robot foundation model, which the startup launched the same week. S1 is designed to learn a previously unseen, long-horizon task from a single video demonstration, using in-context learning rather than task-specific retraining. Nvidia’s post frames the model as a showcase for its physical AI stack: Isaac Lab for reinforcement learning in simulation, Cosmos for turning video into structured training data, Omniverse and Isaac Sim for physically based virtual testing, and TensorRT for optimizing inference once a model reaches a real robot. Nvidia also says it and Skild are jointly building new GPU-accelerated simulation solvers for modeling how robots grip and manipulate objects, due to ship inside Newton for other developers later.

The numbers

Skild AI told Nvidia that S1 succeeded on 66% of steps in new, multistep tasks, against 9% for a comparison AI system, a gap Nvidia calls more than sevenfold. Skild also estimates that one short video demonstration can be worth roughly 380 hands-on training examples, work it says would otherwise take 50 to 100 hours to collect by hand. The company separately cites a $100 million annualized revenue run rate, reached 10 months after its first commercial deployment, alongside deployment work spanning manufacturing, logistics, inspection, security and food preparation. One demonstrated use case has Skild’s dual-arm manipulators assembling Nvidia Blackwell hardware at a Foxconn facility.

What this means

The figures above come from Skild, not from an independent benchmark, so they describe what one vendor reports about its own model rather than a verified industry comparison. What is verifiable is the toolchain: Nvidia is positioning Isaac Lab, Cosmos and TensorRT as infrastructure that a robotics company can adopt end to end, from synthetic data generation through simulation to deployment, without building that pipeline itself. That matters for buyers because it lowers the cost of trying a new robot foundation model vendor, since the simulation and training tooling underneath is shared rather than proprietary to any single robotics company. It also means a customer’s downstream cost structure is tied partly to Nvidia’s hardware and software choices, not only to whichever robot brain sits on top of them.

Sources

• Skild AI Taps NVIDIA Physical AI to Teach Robots New Tasks From a Single Video, Nvidia newsroom, fetched September 16, 2026.